Scorpio 3.0.2

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Daniel Shawul
Posts: 3762
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: Scorpio 3.0.2

Post by Daniel Shawul » Wed Oct 09, 2019 12:42 pm

Hi stefan,

Have you done a re-installation after I added the J13B.2-200 net? When I said use the latest exe, i meant an update even after 3.0.2 -- i know confusing.The old 3.0.2 exe does not have a WDL head support and even if you add wdl_head 1, it will not recognize it. You should get an error message in the log file that says "Unknown command: wdl_head" or something like that in that case i guess.
This is my fault and I have updated the version number of the exe to 3.0.3 with some more changes. For example, in this new upate, though almost never used own scorpio nets format is changed a bit, and if you don't do reinstallation and still have old scorpio nets provided in 3.0.2 they won't work.

You made me curious so I am starting a tournament between 32930 and J13B.2-200 now, so far it doesn't seem all that bad but i will update later with more games:

Code: Select all

GPU0: Score of scorpio32930 vs scorpioJH: 1 - 3 - 4 [0.375] 8
GPU1: Score of scorpio32930 vs scorpioJH: 0 - 0 - 6 [0.500] 6
GPU2: Score of scorpio32930 vs scorpioJH: 1 - 2 - 5 [0.438] 8
GPU3: Score of scorpio32930 vs scorpioJH: 0 - 1 - 5 [0.417] 6
Games:


regards,
Daniel

Joerg Oster
Posts: 691
Joined: Fri Mar 10, 2006 3:29 pm
Location: Germany

Re: Scorpio 3.0.2

Post by Joerg Oster » Wed Oct 09, 2019 2:19 pm

Hi Daniel,

installation of Scorpio 3.0.3 seems to have worked flawlessly.

Code: Select all

C:\Users\xxxxxx\Scorpio\Scorpio-37-14\bin\Windows>scorpio
feature done=0
Number of cores 4 of 4
ht 4194304 X 16 = 64.0 MB
eht 524288 X 8 = 8.0 MB
pht 32768 X 24 = 0.8 MB
treeht 41943040 X 320 = 12800.0 MB
processors [128]
EgbbProbe 4.3 by Daniel Shawul
egbb_cache 4084 X 8216 = 32.0 MB
0 egbbs loaded !
nn_cache 131072 X 1552 = 194.0 MB
Loading neural network : C:\Users\oster\Scorpio\Scorpio-37-14\nets-maddex/net-maddex.uff
With "FLOAT" precision
Loading graph on /gpu:0
Deserialize required 1615929 microseconds.
0. main_input 7168 = 112 8 8
1. value_head 1 = 1 1 1
2. policy_head 1858 = 1858 1 1
Neural network loaded !
loading_time = 2s
uci
id name Scorpio 3.0.3
id author Daniel Shawul
option name log type check default false
option name clear_hash type button
option name resign type spin default 800 min 100 max 30000
option name cores type spin default 1 min 1 max 512
option name ht type spin default 64 min 1 max 131072
option name eht type spin default 8 min 1 max 16384
option name pht type spin default 1 min 1 max 256
option name egbb_path type string default C:\Users\oster\Scorpio\Scorpio-37-14\nnprobe-windows-gpu\
option name nn_path type string default C:\Users\oster\Scorpio\Scorpio-37-14\nets-maddex/net-maddex.uff
option name egbb_cache_size type spin default 32 min 1 max 16384
option name egbb_load_type type spin default 3 min 0 max 3
option name egbb_depth_limit type spin default 6 min 0 max 64
option name egbb_ply_limit_percent type spin default 75 min 0 max 100
option name nn_cache_size type spin default 200 min 1 max 16384
option name n_devices type spin default 1 min 1 max 128
option name device_type type combo default GPU var CPU var GPU
option name delay type spin default 0 min 0 max 1000
option name float_type type combo default FLOAT var FLOAT var HALF var INT8
option name nn_type type spin default 1 min 0 max 10
option name wdl_head type check default false
option name win_weight type spin default 100 min 0 max 1000
option name draw_weight type spin default 100 min 0 max 1000
option name loss_weight type spin default 100 min 0 max 1000
option name smp_type type combo default YBW var YBW var ABDADA var SHT
option name smp_depth type spin default 8 min 1 max 10
option name use_singular type check default true
option name use_probcut type check default false
option name singular_margin type spin default 32 min 0 max 1000
option name probcut_margin type spin default 195 min 0 max 1000
option name aspiration_window type spin default 10 min 0 max 100
option name contempt type spin default 10 min 0 max 100
option name cpuct_init type spin default 125 min 0 max 1000
option name cpuct_base type spin default 19652 min 0 max 100000000
option name policy_temp type spin default 235 min 0 max 1000
option name noise_alpha type spin default 30 min 0 max 100
option name noise_beta type spin default 100 min 0 max 100
option name noise_frac type spin default 25 min 0 max 100
option name noise_ply type spin default 30 min 0 max 100
option name fpu_red type spin default 33 min -1000 max 1000
option name fpu_is_loss type check default false
option name reuse_tree type check default true
option name backup_type type combo default CLASSIC var MINMAX var AVERAGE var MIX var MINMAX_MEM var AVERAGE_MEM var MIX_MEM var CLASSIC var MIX_VISIT
option name frac_alphabeta type spin default 0 min 0 max 100
option name frac_freeze_tree type spin default 100 min 0 max 100
option name frac_abrollouts type spin default 20 min 0 max 100
option name frac_abprior type spin default 0 min 0 max 100
option name mcts_strategy_depth type spin default 30 min 0 max 100
option name alphabeta_depth type spin default 16 min 1 max 100
option name evaluate_depth type spin default 0 min -4 max 100
option name virtual_loss type spin default 1 min 0 max 1000
option name visit_threshold type spin default 800 min 0 max 1000000
option name montecarlo type check default true
option name treeht type spin default 12800 min 0 max 131072
uciok
go depth 12
# rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
# [sd = 12 , hply = 0]
# Move   Value=(V,P,V+P)   Policy  Visits                  PV
# ----------------------------------------------------------------------------------
#  1   (0.527,0.527,0.527)  19.60       0   d2-d4
#  2   (0.527,0.527,0.527)  19.39       0   e2-e4
#  3   (0.527,0.527,0.527)   9.25       0   c2-c4
#  4   (0.527,0.527,0.527)   8.98       0   Ng1-f3
#  5   (0.527,0.527,0.527)   4.83       0   g2-g3
#  6   (0.527,0.527,0.527)   4.49       0   e2-e3
#  7   (0.527,0.527,0.527)   3.50       0   b2-b3
#  8   (0.527,0.527,0.527)   3.31       0   c2-c3
#  9   (0.527,0.527,0.527)   3.16       0   d2-d3
# 10   (0.527,0.527,0.527)   3.15       0   Nb1-c3
# 11   (0.527,0.527,0.527)   2.94       0   f2-f4
# 12   (0.527,0.527,0.527)   2.53       0   b2-b4
# 13   (0.527,0.527,0.527)   2.39       0   h2-h3
# 14   (0.527,0.527,0.527)   2.35       0   a2-a3
# 15   (0.527,0.527,0.527)   2.04       0   Ng1-h3
# 16   (0.527,0.527,0.527)   1.83       0   a2-a4
# 17   (0.527,0.527,0.527)   1.67       0   h2-h4
# 18   (0.527,0.527,0.527)   1.64       0   g2-g4
# 19   (0.527,0.527,0.527)   1.59       0   f2-f3
# 20   (0.527,0.527,0.527)   1.35       0   Nb1-a3
# nodes = 0 <-2147483648% qnodes> time = 399ms nps = 0 eps = 0 nneps = 5
# Tree: nodes = 21 depth = 0 pps = 5 visits = 1
#       qsearch_calls = 0 search_calls = 0
bestmove d2d4
quit
Bye Bye
I have two questions, though.

Do you recommend installing Scorpio as UCI engine?
Can I use the J13B.2-200 net with 'INT8' on my GTX1050Ti?
Jörg Oster

Daniel Shawul
Posts: 3762
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: Scorpio 3.0.2

Post by Daniel Shawul » Wed Oct 09, 2019 2:29 pm

Final score:

Code: Select all

GPU0: Score of scorpio32930 vs scorpioJH: 2 - 7 - 26 [0.429] 35
GPU1: Score of scorpio32930 vs scorpioJH: 1 - 3 - 28 [0.469] 32
GPU2: Score of scorpio32930 vs scorpioJH: 3 - 4 - 26 [0.485] 33
GPU3: Score of scorpio32930 vs scorpioJH: 3 - 3 - 26 [0.500] 32
It has won against 32930 but not by a huge margin. Btw for this test i used HALF precision since my test machine is volta.
I think 42510, that is a 20b net, did better against 32930 so maybe the size is affecting results.

@Joerg

Yes, you can use INT8 on a GTX 1050 Ti and it will be twice faster. Try the 32930 first as that seems to work awesome
even on low end GPU cards with INT8. Then you can play around with J13B.2-200 net and see how it goes.

If you can, please use xboard. If you don't have any other option use uci.

Btw, scorpio will not understand "go depth" command in mcts mode so try "go movetime 20000" .

regards,
Daniel

User avatar
Werner
Posts: 2428
Joined: Wed Mar 08, 2006 9:09 pm

Re: Scorpio 3.0.2

Post by Werner » Wed Oct 09, 2019 3:04 pm

Hi Daniel,
I played too a few games with JH Net and Scorpio 3.02 with wdl_head 1:
Komodo 13.1 x64 1CPU reached 45% in 33 games (-35 Points). Against the Maddex Net I have had 58,8% for Komodo and against LCZero 0.21.2 Cuda (42668) Komodo reached 45% in 100 games. So maybe it works here, but I am not sure.
When I try to analyze your game posted here against 32930 I get more ore less higher evals in the same depth and sometimes other moves e.g.

18...Nc6-e5 0.36/70 instead of ..Rc8 0.33/70
19... Qh6 0.74/71 instead of ...Qh6 0.46/71
(About 3 different moves till move 29)

Here is my Output with Version 3.03:
feature done=0
Number of cores 4 of 4
ht 4194304 X 16 = 64.0 MB
eht 524288 X 8 = 8.0 MB
pht 32768 X 24 = 0.8 MB
treeht 41943040 X 320 = 12800.0 MB
processors [64]
EgbbProbe 4.3 by Daniel Shawul
egbb_cache 4084 X 8216 = 32.0 MB
180 egbbs loaded !
nn_cache 131072 X 1568 = 196.0 MB
Loading neural network : c:\Users\6600K\Arena\Engines\Scorpio302\Scorpio-14-02\nets-lczero\J13B.2-200.uff
With "INT8" precision
Loading graph on /gpu:0
Deserialize required 2181105 microseconds.
0. main_input 7168 = 112 8 8
1. value_head 3 = 3 1 1
2. policy_head 1858 = 1858 1 1
Neural network loaded !
loading_time = 3s
# rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1
# [st = 10835ms, mt = 29250ms , hply = 0 , moves_left 10]
63 26 110 1948 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3
64 28 221 4151 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8
65 30 334 6536 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8 Ke1-g1
66 31 445 8931 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6
67 31 557 11490 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7
68 32 670 14019 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7 Rf1-d1
69 32 785 16678 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 d7-d5 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7 Rf1-d1 b7-b6
70 32 895 19202 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 d7-d5 Bf1-g2 Bf8-b4 Bc1-d2 Bb4-e7 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7
71 33 998 21542 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 d7-d5 Bf1-g2 Bf8-b4 Bc1-d2 Bb4-e7 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7 Rf1-d1
# Move Value=(V,P,V+P) Policy Visits PV
# ----------------------------------------------------------------------------------
# 1 (0.551,0.532,0.541) 18.50 7206 e2-e4 e7-e5 Ng1-f3 Nb8-c6 Bf1-b5 Ng8-f6 Ke1-g1 Nf6xe4 d2-d4 Ne4-d6 Bb5xc6 d7xc6 d4xe5 Nd6-f5 Qd1xd8 Ke8xd8 h2-h3 Bf8-e7 Nb1-c3 h7-h5 Bc1-f4 Kd8-e8 Ra1-d1
# 2 (0.553,0.532,0.542) 17.16 9840 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 d7-d5 Bf1-g2 Bf8-b4 Bc1-d2 Bb4-e7 Ng1-f3 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7 Rf1-d1
# 3 (0.541,0.532,0.536) 8.85 1191 Ng1-f3 d7-d5 d2-d4 Ng8-f6 c2-c4 e7-e6 g2-g3 Bf8-b4 Bc1-d2 Bb4-e7 Bf1-g2 Ke8-g8 Ke1-g1 c7-c6 Qd1-c2 Nb8-d7
# 4 (0.545,0.532,0.538) 8.25 1286 c2-c4 e7-e5 g2-g3 Ng8-f6 Bf1-g2 Bf8-c5 d2-d3 Nb8-c6 Nb1-c3 Ke8-g8 a2-a3 a7-a5 e2-e3 Bc5-a7 Ra1-b1 h7-h6 b2-b4 a5xb4 a3xb4 d7-d6
# 5 (0.537,0.532,0.534) 5.60 552 e2-e3 Ng8-f6 Ng1-f3 c7-c5 d2-d4 d7-d5 Bf1-e2 Nb8-c6 c2-c4 d5xc4 Be2xc4
# 6 (0.526,0.532,0.526) 5.01 324 g2-g3 d7-d5 Ng1-f3 c7-c5 Bf1-g2 e7-e6 Ke1-g1 Ng8-f6 c2-c4 d5-d4 d2-d3 Nb8-c6 e2-e3
# 7 (0.531,0.532,0.531) 3.68 278 c2-c3 Ng8-f6 d2-d4 d7-d5 Ng1-f3 e7-e6 Bc1-f4 Bf8-d6 e2-e3 c7-c5 Bf4-g3 Ke8-g8
# 8 (0.520,0.532,0.520) 3.25 171 Nb1-c3 d7-d5 d2-d4 Ng8-f6 Bc1-f4 a7-a6 e2-e3 e7-e6 g2-g4 Bf8-b4 Ng1-e2 Ke8-g8 a2-a3 Bb4-a5
# 9 (0.507,0.532,0.507) 3.10 119 d2-d3 d7-d5 Ng1-f3 Ng8-f6 g2-g3 Nb8-c6 d3-d4 Bc8-f5 Bf1-g2 e7-e6 Ke1-g1 Bf8-e7 c2-c4 Ke8-g8
# 10 (0.503,0.532,0.503) 3.09 110 b2-b3 e7-e5 Bc1-b2 Nb8-c6 e2-e3 Ng8-f6 Ng1-f3 e5-e4 Nf3-d4 Bf8-c5 d2-d3 Nc6xd4 e3xd4 Bc5-b4 c2-c3 Bb4-e7 d3xe4 Nf6xe4 Bf1-d3 d7-d5 Nb1-d2 Ne4xd2 Qd1xd2
# 11 (0.506,0.532,0.506) 3.05 115 a2-a3 d7-d5 d2-d4 Ng8-f6 e2-e3 Bc8-f5 c2-c4 e7-e6 Qd1-b3
# 12 (0.486,0.532,0.486) 2.95 80 a2-a4 Ng8-f6 Ng1-f3 c7-c5 e2-e3 Nb8-c6 d2-d4 d7-d5 c2-c3
# 13 (0.490,0.532,0.490) 2.69 76 h2-h3 e7-e5 e2-e4 Ng8-f6 Nb1-c3 Bf8-b4 a2-a3 Bb4-a5 Ng1-f3 Ke8-g8 Bf1-c4
# 14 (0.443,0.532,0.443) 2.68 45 f2-f4 d7-d5 Ng1-f3 c7-c5 e2-e3 Ng8-f6 Bf1-b5 Bc8-d7 Qd1-e2
# 15 (0.462,0.532,0.462) 2.31 45 b2-b4 e7-e5 Bc1-b2 Bf8xb4 Bb2xe5 Ng8-f6 Ng1-f3 Ke8-g8 c2-c4 d7-d5
# 16 (0.436,0.532,0.436) 2.11 32 Nb1-a3 e7-e5 Na3-c4 Nb8-c6 e2-e4 d7-d5 e4xd5 Qd8xd5 Nc4-e3
# 17 (0.417,0.532,0.417) 2.09 27 Ng1-h3 d7-d5 g2-g3 c7-c5 Bf1-g2 Nb8-c6 d2-d3
# 18 (0.379,0.532,0.379) 2.01 20 f2-f3 e7-e5 e2-e4 Nb8-c6 Nb1-c3 Bf8-c5 f3-f4
# 19 (0.425,0.532,0.425) 2.00 27 h2-h4 e7-e5 c2-c4 Ng8-f6 g2-g3 Bf8-c5 Ng1-f3 e5-e4 d2-d4 Bc5-b4 Bc1-d2
# 20 (0.271,0.532,0.271) 1.62 10 g2-g4 d7-d5 h2-h3 h7-h5 g4xh5 e7-e5 d2-d3 Nb8-c6
# nodes = 215738 <0% qnodes> time = 10263ms nps = 21020 eps = 0 nneps = 2123
# Tree: nodes = 28267 depth = 29 pps = 2113 visits = 21555
# qsearch_calls = 0 search_calls = 0
move d2d4

I have made a new .trt file - and it was different to that of Version 3.0.2 with wdl_head 1 (do the other new Settings make the difference or is this file every time different ??

PS. do you think Scorpio works on Windows 7 too ?

best wishes
Werner
Werner

Daniel Shawul
Posts: 3762
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: Scorpio 3.0.2

Post by Daniel Shawul » Wed Oct 09, 2019 3:17 pm

Hi Werner,
I you use 3.0.3 it should work fine -- the scorpio.ini should already have a wdl_head entry in it.
Since the UFF file is optimized for different GPUs, especially with INT8, it is not going to be
exactly the same for everybody. So I wouldn't worry about slight differences, and your run with 3.0.3 seem fine for me.

The previous problem was that when I said JH net is added, I should also have mentioned it needed reinstallation or
redownloading and updating the scorpio exe.

It should work on windows 7 too but the installation requires powershell which may not come with windows 7.

The *.trt file changes only when you update the TensorRT package (which i haven't). For each combination of threads (mt) and precision(FLOAT/HALF/INT8), it will generate a separate .trt file.

Daniel

pohl4711
Posts: 1119
Joined: Sat Sep 03, 2011 5:25 am
Location: Berlin, Germany
Contact:

Re: Scorpio 3.0.2

Post by pohl4711 » Wed Oct 09, 2019 3:42 pm

Daniel Shawul wrote:
Wed Oct 09, 2019 12:42 pm
Hi stefan,

Have you done a re-installation after I added the J13B.2-200 net? When I said use the latest exe, i meant an update even after 3.0.2 -- i know confusing.The old 3.0.2 exe does not have a WDL head support and even if you add wdl_head 1, it will not recognize it.
Of course not. How should I know?!

I will continue the testrun of 32930 until it is complete, now. I cannot restart my testruns over and over again.

User avatar
Werner
Posts: 2428
Joined: Wed Mar 08, 2006 9:09 pm

Re: Scorpio 3.0.2

Post by Werner » Sat Oct 12, 2019 12:27 pm

Here a match on my GTX 1650 with J13B.2-200 net :

1 Scorpio 3.0.3 MCTS+NN J13B.2-200 +53 +29/=57/-14 57.50% 57.5/100
2 Komodo 13.1 x64 1CPU -53 +14/=57/-29 42.50% 42.5/100

In comparison with LCZero and NN 42668 it scores a bit better:

1 LCZero 0.21.2 (42668) +28 +31/=46/-23 54.00% 54.0/100
2 Komodo 13.1 x64 1CPU -28 +23/=46/-31 46.00% 46.0/100

so I think, this network runs correct with Scorpio.
Werner

Daniel Shawul
Posts: 3762
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: Scorpio 3.0.2

Post by Daniel Shawul » Sat Oct 12, 2019 3:24 pm

Hi Werner,

Thanks for the test ! I think so too and so far my tests also put it slightly above 32930.
If it works this good on a GTX 1650, i think bigger nets are the way to go with upcoming nvidia gpus.

Btw I gave up on trying to get T40 nets to work with INT8 -- i feel like there is something different the way it was trained
that breaks INT8 because other nets with almost identical architecture, such as MaddexSE, work just fine. Even the latest
advanced network architecture with covolution policy + wdl head + SE work fine but not T40.

Daniel

User avatar
Werner
Posts: 2428
Joined: Wed Mar 08, 2006 9:09 pm

Re: Scorpio 3.0.2

Post by Werner » Sat Oct 12, 2019 5:26 pm

Hi Daniel,
that´s a pity - I thought perhaps J.Horthos can make a good net for you based on maddex or your own nets - which is much faster and will not need another year :-) ;

And of course the J13B nets are a bit too large for my GTX 1650 - but in comparison to my CPUs I need no faster GPU as I want to reach a ratio between 1 and 1.3 - which I have now (with J13 I am now around 0,9-1,0) - perhaps the upcoming GTX 1650 super (with 75 watt) will do it.

PS did you update Scorpio ? - I saw you found a bug...
Werner

User avatar
Werner
Posts: 2428
Joined: Wed Mar 08, 2006 9:09 pm

Re: Scorpio 3.0.2

Post by Werner » Sun Oct 13, 2019 4:08 pm

After 32 games now I am not so sure if Scorpio with J13 net runs correct here:

LC0 0.22.0 J13B.2-200 - Komodo 13.1 x64 1CPU 22-10

Scorpio result was 18-14

I will continue the match.
Werner

Post Reply